扁平化办公白领职场人士场景讨论元素

生成一张TASLP Introduction 用的三栏逻辑图（横向1×3，标(a)(b)(c)）。要求白底极简、文字逻辑为主，只用少量符号(→, ×, ✓, 🔒)，尽量少icon；每栏用矩形框+短句，栏与栏之间用粗箭头连接表示“因此/于是”。(a) 标题：(a) Centralized training: privacy problem内容用3行：Train ADD needs large real speechCentralized pipeline uploads raw audioRaw audio contains identity/content 🔒 → leakage/compliance risk ×底部小字：Raw audio cannot leave devices(b) 标题：(b) Federated learning: necessary but challenging3行：Use FL: data stays local 🔒Upload only model updates (weights/gradients)Real-world FL: Non-IID + dropouts/partial participation → performance gap (EER↑), instability ×(c) 标题：(c) Our FLADD: close the performance gap under privacy4行（只点名用途，不画细节）：Goal: high performance under privacy constraintDual-branch detector → stronger representationDual prototypes (P1/P2) → align heterogeneous clientsMFAW aggregation → robust & stable training右下角加：Outcome: high-accuracy & stable ADD ✓ with no raw-audio sharing 🔒整体底部加一句：Privacy is a hard constraint, not an add